Deep Reinforcement Learning Trajectory Control of Nonholonomic Autonomous Surface Vehicles in Challenging Wave Conditions: Theory and Preliminary Evaluation

Autonomous Surface Vehicles (ASVs) are vital in maritime operations, requiring precise trajectory tracking, with rough sea conditions as a challenge. Proportional Integrative and Derivative (PID) and Model Predictive Control (MPC) controllers are widely used, yet PID controllers face challenges with coupled motions sum to nonlinearities, and MPC requires an accurate system dynamics model that frequently is unavailable or requires inaccessible resources. This paper proposes a model-free Deep Reinforcement Learning (DRL) approach for trajectory tracking under rough sea conditions without relying on system dynamics models and capable of controlling both holonomic and nonholonomic ASV configurations. Nonholonomic configurations have the advantage of requiring fewer thrusters, thus being more efficient. We compare state-of-the-art model-free controllers with our Deep Reinforcement Learning (DRL) method, evaluating performance based on RMSE for position and heading using holonomic and nonholonomic ASV configurations with four and two thrusters, respectively. Results, validated with simulated data, confirm the effectiveness of our DRL method.